Rise of the citizen developer: GenAI and the democratisation of code

Computer Weekly asks artificial intelligence and data specialists for their take on the large language model-powered rise of so-called ‘citizen developers’

App development harnessing generative artificial intelligence (GenAI) might be democratising coding and potentially freeing up resource, but organisations must carefully manage contributors without prior coding experience.

According to Jon Puleston, vice-president of innovation in the profiles division at Kantar, even an accurate synthetic persona based on a data set of 25,000 real people containing 250 predictive demographic variables can give “very erratic” results even to simple questions, such as ‘Do you own a dog?’.

“In our experiment, this could only be predicted with 75% accuracy even using the most advanced AI and machine learning [ML] technique we applied, against a 52% chance based on our demographic data,” Puleston says. “This highlights the sheer amount of inputs required to make synthetic data models accurate enough for generalised commercial use. When making a business decision worth billions of dollars, there are risks.”

Kantar’s experiment suggests complexities count against talk of “citizen developers” when it comes to the rise of natural-language generative-AI tools in particular, with Puleston adding: “Real human insights are still the heart of good market research.”

After all, the detailed context of any given activity can often prove crucial. Kicking a football between goalposts is in principle a simple act, but it takes years of practice to reliably do well.

Leslie Kanthan, CEO and co-founder of code optimisation company TurinTech, says that traps await the unwary – even if an organisation has its own datasets ready to go. “Engaging with AI code and in-house applications is good because it gets people using this cutting edge stuff, allowing idea generation and creativity. But there’s the governance of that,” Kanthan says.

If creating an application with sensitive data sets, for instance, without a coding background, you’ll be hard-pressed to justify to regulators and compliance exactly what the application is doing and not doing.

“Just saying that you built an application and it solves this problem is not good enough for regulators,” Kanthan says. “They need to know that the data is being kept right, and so forth.”

That said, GenAI can help junior developers a year or two in sharpen their edge, and development might be expanded to wider contributions while remembering the need for robust testing, review and user adoption as tools advance.

Do not forget that cleaner code can be more sustainable, consuming less resource including energy, Kanthan adds. “Right now, think about [GenAI] realistically. It’ll be a few years still before you have robust applicational use for even your most advanced uses,” he says. “You need to know exactly what you’re doing.”

Ask what exactly you are bringing into the company and how to make use of it. Today’s GenAI is best used as an assistant to expertise, not a substitute for it – and develop appropriate checks and balances and governance in every use case. “Know what you’ve created and how you’ve been able to justify that,” Kanthan adds.

Kjell Carlsson, head of AI strategy at Domino Data Lab, notes that offerings such as Amazon Q billed as accessible to non-coders will typically be of limited use. The question is not whether you need skills to use a tool, but can you actually use that tool to good effect? A rule of thumb is conversely that you’ll need to “up your humans”.

Compare the challenge with using Visual Basic or Excel macros, for instance: those take skill or a technical background. And what about documentation?

In businesses with less resource, where less-experienced developers might adopt ChatGPT, for instance, to build apps, you cannot really have “a whole lot” of confidence in the quality of the results, Carlsson confirms.

“It takes a lot of understanding on the part of an end user to know what good looks like,” he says. “It can be a great way to get up to speed quickly and do bog-standard things. Not if it’s a task that you usually use SQL for or – heaven forbid – COBOL.”

‘Anyone can code’ 

Gavin Harcourt, engineering lead at marketing intelligence platform Streetbees, says their experience with GenAI, working through product design, discovery and early implementation, highlighted the myth that “anyone can code”.

“With these demos of, say, Amazon Q, I have a lot of scepticism: we learned you have to invest,” Harcourt says. “We have teams of people schematising, physically translating market research expertise and more.”

Shaf Shajahan, Streetbees’ AI product director, adds: “That’s including two-and-a-half head-count dedicated solely to translating complex market research problems into data libraries, like strategy consultants learning how to navigate complex data schemas. You need marketing, strategy and tech together, and people dedicated to engineering.”

GenAI will evolve, with large language models (LLMs) becoming more commoditised over time. Yet, even with “an element of quality and output” that almost anyone can deliver, that’s probably not optimum, Shajahan notes.

“Your key differentiator will be that it’s not just a thin wrapper around a core LLM, it’s through the quality that your product or your output, that it has been designed and developed by experts in the domain,” Shajahan notes.

Hans de Visser, chief product officer (CPO) at low-code developer platform vendor Mendix, says they divide AI-assisted development, which is about inserting AI services for its devs working in its platform to boost their productivity, from AI-augmented applications that weave AI services into smarter applications built by the developer.

Policies, practice and governance specific to requirement are paramount. Some customers seek “shift left” partly with a view to reducing workloads on the more technical teams and giving workers more control over their own destiny. “But think through the type of application you’re talking about, and what platform is suitable to support that,” Visser says.

Examine scope, reach and criticality of individual work groups, departments and the enterprise versus application complexity. If you generate a visual domain or data model that you can completely see through, you should then be able to validate results and whether they apply for the relevant page or workflow generated.

Think about the “walled garden” of Microsoft PowerApps, which in its simplest form can have SharePoint and Excel as mere sources for connecting data. Around that, you can build a user interface, a checklist or very simple workflow for approval reminders or similar.

“If you keep it in that boundary and can govern it, you might be totally fine for citizen developers and using the kinds of AI features in that particular platform,” Visser says.

For more sophisticated departmental applications, incorporating GenAI services will boost productivity anyway via abstraction and automation.

“Consider what type of GenAI features you would grant to what type of developer. If you have GenAI based upon code, you will get generated code. If they’ve never built an application they typically can’t make sense of what’s happening behind the scenes,” notes Visser.

Melvyn White, principal enterprise architect for AI/ML at BT Group, broadly agrees with the need to match so-called democratisation with development restricted to a certain level of expertise – pending rapid technological evolution and organisations’ need to keep pace.

“Opening AI up for the masses is in many ways a good thing,” White says, noting that in some ways GenAI can be thought of as the next step on from using unified modelling language (UML) as an accelerator.

“My friends who program every day use LLMs, but they’re very good programmers in their own right. It does enhance their productivity to some extent, but they still have to dip in and make tweaks and changes. They know what they’re doing.”

Read more about generative AI

Read more on Artificial intelligence, automation and robotics

Data Center
Data Management